Mistral previews a 1-trillion-parameter model, and Claude Haiku 5.5 costs 90% less

Two things moved the frontier this week: a European flagship still behind a preview, and a price cut that shipped to everyone.

Mistral previewed Mistral Large 4 on October 6, 2026: a 1-trillion-parameter, natively multimodal model that the company says will ship open weights by the end of October. Until then it runs as a guardrailed preview on Mistral's own API, red-teamed with cybersecurity partners and state authorities. 1

Artificial Analysis scores it 38 on its Intelligence Index, which makes it the most intelligent model from outside the US and China — level with GPT-6 Luna at its maximum effort setting and with DeepSeek V4.1 Flash. At $1.13 per index task on standard pricing, Mistral Large 4 costs over four times as much as GLM-5.3-Flash ($0.25) and DeepSeek V4.1 Flash ($0.27), models of similar measured intelligence. 2

On one Cyber Index test — reproduce a real vulnerability in open-source software, then patch it — Mistral Large 4 scores 82%, the highest of any model. Claude Opus 5.5 and GPT-6 Astra score near zero on the same task because they refuse to perform it. 12

Artificial Analysis then opened its Cyber Index to trusted-access models on October 8, 2026. The first one added, GPT-6 Sol (Daybreak Blue), now leads the board, records no safety blocks across the index and improves 32 points on the publicly available GPT-6 Sol, at $1.77 per task. It is reachable only through OpenAI's Daybreak program. 3

Anthropic released Claude Haiku 5.5 on October 7, 2026, at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, with prices rising fivefold above that. Anthropic puts it 90% below Haiku 4.5 on the shorter prompts that accounted for 90% of its predecessor's requests. 4

Artificial Analysis measures Haiku 5.5 at 43 on the Intelligence Index, 26 points above the last Haiku a year earlier. It also flags a defect: a safety refusal issue during pre-release testing made the model over-refuse, holding its AutomationBench score at 35% against 53–60% for models of the same class. 5